{"path":"research/quality-control-for-deeper-steps.md","content":"# Quality control for deeper steps: what the small tasks are held to today, and what writing under a machine's own sentence adds\n\n**Date**: 2026-10-10 · **Type**: a founder ruling, grounded in what the corpus and the code hold\nabout how the small machine tasks are checked, with three counts on the live graph and one small\ntrial of a second level · **Status**: the ruling is his. Every reading and proposal here is\nClaude's and **unruled** (ledger AG7 to AG11). Two things were built from it the same day: the\npage where a person reads what the machine wrote (§ 5.4), and the storing of a machine's sentences\nas claims, on his ruling of that evening (ledger AG12; § 4.2). None of the controls proposed here\nruns on a live task.\n\n**His words**, 2026-10-10, asked whether below its first level the machine may write deeper steps\nitself or should only find and ask until its error is measured (ledger AG6): *\"It may and should\nwrite deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just\nlike the other microtasks that are in play in Deliberus. Ground deeply on this. Use the corpus\ngraph to the best of your ability to make sure we keep track of all connections both known and\nunknown to the rest of the ideas in the corpus.\"*\n\nClaude's pick had been the opposite (only find and ask), and it is withdrawn. His sentences of\n29 September and 7 October already implied his answer (§ 6).\n\n**How it was read.** The corpus graph was asked first, about 42 ledger rows and notes\n(`scripts/corpus_graph.py around …`). Three read-only readers then went through the corpus, each\nwith one part of the question, and their reports are kept word for word beside this note (§ 11).\nA finding of theirs is used here only where both passages were opened afterwards. The notes on\nproving a pass, the trial of 7 October and the code of each small task were read directly. Three\ncounts were taken on the live graph with no model. One trial was run, with its predictions written\nfirst (§ 5).\n\n**Words used here.** A *made* sentence is one the machine wrote in its own words. A *found*\nsentence is one a text states, which the machine pointed at. A *level* is what is written under\none sentence to say why it holds. The *descent* is the pass going down level by level.\n\n---\n\n## 0. In short\n\n1. **The control he points at is real, and its human half has not begun.** Ten controls can be\n   named. Four run on live tasks today: the question is asked small and closed, something other\n   than the writer checks the answer, what fails is held and what lands is marked and can be\n   undone, and every call is on record. One, asking a second way, exists in trial scripts only\n   and has no ruling of his. Five wait on a person. ⚠ Corrected 2026-10-10: this line said *five\n   run today* and counted asking a second way among them, against the table in § 1.1. He caught\n   it: *\"'Asked twice'? I don't think I ever blessed that particular part.\"* No person has yet\n   answered a sitting, written a note on a machine's output, or touched a machine-cut part. § 1.\n2. **So no small task has a measured error rate.** Every figure in the corpus about these tasks is\n   a model reading a model. § 1.2.\n3. **A deeper step is a different kind of write from a split.** The splitter may write what it\n   found in a claim and must hold what it added. A step under a link is added by definition. It can\n   be written because it is filed elsewhere: as the machine's question under a link, part of\n   nobody's claim, weighing nothing. § 2.\n4. **Depth adds five things the first level did not have**: errors that build on each other; a\n   change in the kind of step, into kinds that have no check; checks that agree because they are\n   one family; chains that look sound a step at a time; and a rule's error carried to every link\n   that uses it. § 3.\n5. **The main brake is looking a made sentence up before building under it.** A sentence a text\n   states starts the count again. The worst error available is marking a made sentence as found,\n   and the answer keys to test that exist today, with no person needed: 1,789 claims carry their\n   own quoted passage. § 4.1.\n6. **A wider rule is not a step down.** Asked for the wider sentence a rule is a case of, the\n   writer wrote one for all nineteen rules, twelve of them rules it had twice read as resting on\n   something else. What it writes is a wider wording of the same rule, no easier to check. The way\n   down runs through how a thing comes about and how it is known. § 5.\n7. **The first-level check does not carry over unchanged.** Handed the wider sentence alone, the\n   reader still reached the rule on 13 of 19, and with any one stage of a chain left out it\n   reached the rule every time (36 of 36). At the first level the reader chose among four claims\n   of one text. Under a made rule there is no text to take same-subject alternatives from. § 5.\n8. **The writer's own no can fire, when the form asks for something concrete.** Asked for a chain\n   of what brings about what, it declined two rules of ten and said why. Asked for a wider\n   sentence, it declined none of nineteen. § 5.\n9. **Depth needs brakes that are not the writer's.** The stop rule of 25 August is a weak brake on\n   fresh material: of 1,198 open questions on claims nobody has examined it prices 595 as worth\n   asking, and under the 930 links that carry no pattern it has nothing to read. § 4.7.\n10. **The limit on depth is to be measured by level.** Until it is, the pass writes as deep as its\n    checks pass, every sentence carrying its count, unseen by default. § 4.8.\n\n---\n\n## 1. What his sentence points at: the control the small tasks are under today\n\n*\"Just like the other microtasks that are in play\"* names a regime. His own statement of it is\nfrom 2026-09-21, when he first asked for a machine pass to write by itself: *\"I very much think\naccuracy in this way can be increased enough to be relied upon if we make sure AI/machine use is\ntargeted to scoped/narrow enough passes/microtasks that we can develop trust in over time … We can\nalways mark machine-split claims in some subtle way in order to know their provenance in order to\nreview the accuracy continually … This general approach I also want applied to other\nmachine-vs-human tasks and ratification-dependent processes/workflows etc.\"*\n([earning-the-right-to-write.md](earning-the-right-to-write.md)). And on the order of work, the\nsame day: *\"Wire and switch first, sitting later.\"*\n\n### 1.1 Ten controls, and where each stands\n\n| | Control | Built | Measured | Ruled, in his words | Still a proposal |\n|---|---|---|---|---|---|\n| 1 | **The question is asked small and closed**: one judgement to a call, a refusal asked first as a gate, the reasoning before the verdict | a test that reads every reply form; the gate in the linking step and in the reader of a link's bearing; the gate in the pattern classifier, built 2026-10-10 and not yet live | the pattern classifier's *none fits* fired 0 of 3,666 times beside a costless shrug | *\"Great, do it, and document that rule\"*; *\"Single exit is blessed\"* | |\n| 2 | **It is asked a second way before anything rests on it**: two wordings, both orders | in trial scripts only; the sameness check can read a pair in both orders and the linking step calls it without that, so it runs on no live task | 4 of 30 links refused under one wording, 20 under another; a check moved with the order of two premises on 14 of 44; the first-level reader answered the same in both orders on 26 of 29 | none. Founder, 2026-10-10: *\"'Asked twice'? I don't think I ever blessed that particular part.\"* | AD14; and the note in `CLAUDE.md` that states it as a rule was a session's, added 2026-10-07 |\n| 3 | **Something other than the writer checks it, and code checks first** | the splitter's separate call that checks each part; the sameness check ahead of the pattern classifier; a script that checks quoted passages | the part question answered the same twice on 45 of 45; of five checkers on one split, each had a miss or a false alarm; 94.1% of 1,144 passages stand word for word | *\"Should we rather not just fix the \"coverage check\"?\"* | Y4, Y6 |\n| 4 | **What fails is held; what lands is marked, can be undone and weighs little** | the splitter holds a failed cut; a written part carries its writer, version and check; a replaced link stays visible; a link the machine guessed presumes no bearing | the first splitter held 4 of 11 for invention; 6 of its 37 written parts stated as fact what their claim denied or conceded, found by another check | *\"Wire and switch first, sitting later.\"*; *\"Yes, replace by itself, keep the old link visible.\"* | AF2 |\n| 5 | **Every call is on record**: model that answered, instructions, inputs, reply; the fingerprint of the instructions is the version | since 2026-10-06; the database refuses a change or a deletion of a row | every call of this note's own trial is in it, the retried ones too | *\"Yes build it.\"* | the record's number on each written item (Z1b) |\n| 6 | **A person reads first, and their touch is the measure** | two sittings, one with planted wrong pairs (1 in 6) and both with 1 item in 5 shown blind; a note box on every verdict; keep, reword or drop on a machine-cut part | **nothing**: § 1.2 | *\"Sure\"*, to reading before proving; *\"Add it!\"*, to the note box | when a first reading ends (Y9b); a sitting for the passes that write (Y10b); a weekly sitting (Y14) |\n| 7 | **A rate is proven per version and per kind**: blind, random, about 60 clean items for under 5 in 100 and about 150 for under 2 | the arithmetic | no pass is proven | | T3, Y1, Y11, Y12 |\n| 8 | **What a pass may do follows what is proven**, and it drops a rung when its errors rise | the splitter writes marked, by his ruling | | *\"Wire and switch first, sitting later.\"* | the ladder itself, and the drop: which writers are presumed is a fixed list in the code |\n| 9 | **Writer and checker from different model families** | the answering model is stored by the splitter and by the reader of bearing | on one real split, three checkers of the writer's family passed a fault and one of another family caught it | *\"your proposal is blessed, make it so.\"*, for the reader of a link's bearing | Y4 |\n| 10 | **Machine work built on machine work is capped by measured error** | nothing counts depth; one guard: a claim that has parts is not cut again | no rate, so no cap | *\"Yes.\"* (29 September) | how depth is counted (T1's second half) |\n\nWhere each fact stands in the files: the first reader's report (§ 11), with file and line.\n\n### 1.2 What a person has read so far: nothing\n\nCounted on 2026-10-10 ([the script](quality-control-for-deeper-steps/what_people_have_read.py)).\n\n| | |\n|---|---|\n| Answers given in the sameness sitting | 0 |\n| Answers given in the review of machine-cut parts | 0 |\n| Notes written on any machine verdict (`scripts/read_review_notes.py`) | none |\n| Machine-cut part links, and those a person has kept, reworded or dropped | 111, and 0 |\n\n**So *quality-controlled like the other small tasks* means controls 1 to 5 today.** They are habits\nof asking and marking. They make a wrong answer rarer, visible and undoable. They do not say how\noften an answer is wrong. No small task here has an error rate measured against a person, and the\nproofs and the ladder of controls 7 and 8 have nothing to stand on until control 6 has happened once.\n\nHis end-state of 2026-08-31 says the same from the other side: nothing waits on a person's\napproval, machine structure carries its provenance for good, and what people do with it is the\nmeasurement ([the-ratification-record-and-the-triage-feed.md](the-ratification-record-and-the-triage-feed.md) § 4b).\n\n### 1.3 Five holes found in the regime while reading it\n\nEach was a reader's finding, opened afterwards in the code, and the last was then measured. None\nchanges the design below except the last. All five are in `TODO.md`.\n\n1. **The splitter's separate check is separate, and it is not blind.** It is one call. Its prompt\n   says *\"You have NOT been shown the original claim\"* and prints the original a few lines further\n   down (`deliberus/extraction/decompose_claim.py`). So the sentence it rebuilds from the parts is\n   written by a model that has the original in front of it. Five places said otherwise and are\n   corrected with this note: the module's own header, two always-loaded lines that called the\n   rebuilding unanchored, and two notes. The check of each part needs the original and is\n   unaffected. Making the rebuilding blind means a call of its own, which changes the pass's\n   version, so it is a to-do and not done here.\n2. **The sameness check's order control is built and switched off where linking calls it**, and a\n   *same* verdict marks a pair whether or not the check called itself confident.\n3. **The built reader of a link's bearing passes over all 822 links the machine drew.** It selects\n   links that carry a pattern and are not marked unconfirmed. Those links carry none and are marked.\n4. **Dropping a rung is not built.** Which writers are presumed is a fixed list.\n5. **The stop rule is a weak brake on fresh material** (§ 4.7).\n\n---\n\n## 2. What kind of write a deeper step is\n\n**The splitter's rule and his ruling look opposed, and are not.** The splitter may write what it\n*found* in a claim and must hold what it *added*: a part earns its place because it was already in\nthe sentence, and a part the sentence never asserted would be put in its author's mouth and would\ncap the claim ([earning-the-right-to-write.md](earning-the-right-to-write.md) § 4a). A step under a\nlink is added by definition. On 7 October thirteen of fifteen written rules stood in neither text.\n\n**It can still be written, because it is filed somewhere else.** *(Since his ruling of that\nevening, ledger AG12, a made sentence is a claim of its own that nobody asserts, kept as a part of\nthe link's relevance claim. What follows still holds: it is no part of an author's claim.)* A made\nsentence is never a part of anybody's claim. It is the machine's sentence, attributed to the machine, shown as a question on\nthe link with a move beside it, weighing nothing and capping nothing (ledger AD17). What the\nsplitter's rule protects, an author's claim and the strength of a whole, a made sentence cannot\ntouch.\n\n**What a wrong made sentence can do**, then:\n\n- mislead a reader who opens it. A wrong explanation left readers worse off than none in one\n  study the corpus holds, and a machine's premise can make a weak step look reasoned;\n- waste what is written under it;\n- be marked as found when it is not, and so borrow a text's standing (§ 4.1);\n- tilt, across many links, toward whatever the machine finds easy to say.\n\n**What it cannot do, as long as one condition holds**: move a strength, cap a claim, or reach a\nreader as something an author said. The condition is that nothing reads a made sentence except\nthe descent and its checks (§ 4.6).\n\n**Which quality level applies.** The open row on quality levels (ledger Y1) allows 5 to 10 wrong\nin 100 for a marked, undoable item that moves no strength *and that no later machine work stands\non*. A made sentence with a made sentence under it fails the last clause. So the bar for deeper\nsteps is set on the chain: how likely it is that a sentence at a given depth stands on a wrong one\n(§ 4.8).\n\n---\n\n## 3. What depth adds\n\n### 3.1 Errors build on each other\n\n**This is his ruling of 29 September.** A pass whose writes nothing builds on keeps the same share\nof wrong content at any volume. A step built on a step inherits the error above it. At 2 wrong in\n100 a step, two steps are wrong about 4 times in 100, five about 10, ten about 18\n([proving-a-pass-reliable.md](proving-a-pass-reliable.md) § 1). That arithmetic assumes the errors\nare independent, and they are not: outside measurements find that a model's error rate rises when\nits context holds its own earlier errors, and that taking earlier output out of the context helps\n(a helper's reading of 7 October, unverified:\n[the report](writing-out-the-unwritten/outside-reading-dependable-small-tasks.md)).\n\n**The count starts earlier than it looks.** A text is read into claims: one machine step. The\nlinking step draws a link: two. A rule is written under it: three. Something is written under the\nrule: four. The second level under a machine-drawn link is the fourth machine step from the text,\nand none of the four has a measured rate.\n\n### 3.2 One level down, the kind of step changes\n\n**Four shapes of step are on record, and one has a check** (ledger AD23): from a rule and a case;\nfrom a sign to what it indicates; from a case to a like case; from cases to a rule. The three-part\nform and the reader of 7 October fit the first only.\n\n**What sits under a rule is mostly not the first shape.** In the trial of § 5, nineteen\nmachine-written rules were asked what kind of answer *why does this hold* calls for, twice. Read\nthe same both times: how something works, 6; a wider rule, 4; what its words mean, 4; unsteady, 5.\nSo for most rules the next level is a kind of step the built check was not made for.\n\n### 3.3 Checks from one family agree with each other\n\nA step that passes three checks from one family of models has been checked fewer than three times.\nThe corpus holds this from four sides: the reader of 7 October joined two names silently because\nthe writer does; on one real split three checkers of the writer's family passed a fault that a\nmodel of another family caught; nine judges from seven makers carried about two independent votes\nin one outside measurement; and one outside analysis finds that with correlated checkers failure\nfalls off far more slowly than multiplying their rates suggests, with a share of errors no number\nof checks reaches (helper readings, unverified). A wrong step is rejected about half the time by a\nmodel judging one step, in the one outside figure the corpus has.\n\n### 3.4 A chain can be sound a step at a time and wrong as a whole\n\nFour ways, each already named in the corpus.\n\n- **Support does not pass along a chain in general.** It does when the middle claim screens the\n  first off from the last. So each hop is owed its own first half: does the claim above support\n  the machine's sentence ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 7).\n- **A chain that mixes directions can manufacture support.** A step from a sign to its cause and a\n  step from a cause to its effect do not chain freely (Pearl 1988, an abstract read by a helper).\n  His squirrel is such a chain: a sign step, then causal steps under its law.\n- **Error can be spread thin** across many steps that each look plausible, so that it is nowhere in\n  particular. The threat model has held this since July, and its named guard is that a\n  decomposition come out the same when it is made again. A first form of that guard exists and\n  refiles by code; no form of it makes a decomposition a second time (§ 9).\n- **A chain can drift or circle**: by its third level it is about something else, or it says its\n  first sentence over again. A prediction of this was written down on 29 September and has had\n  nothing to be tested on: *drift compounds with each hop, so there is a depth beyond which the\n  leaves no longer rebuild the root; refuted if coverage holds flat as depth grows*\n  ([major-transitions-heredity-and-evolvability.md](major-transitions-heredity-and-evolvability.md)).\n  A pass that writes deeper steps is the first thing that can test it.\n\n### 3.5 A rule used under many links carries its error to each\n\nA made rule is stored once and counted across the links that rest on it (ledger AD18). Its count\nis where attention pays. It is also how far a wrong rule reaches. So the checking of a rule is\nweighed by how many links use it.\n\n### 3.6 Depth has been a sign of care\n\nThe public page on depth says a descent's depth at any point is how far anyone has cared to go,\nand that this unevenness is the information ([depth.md](../depth.md)). A machine that fills every\nlevel makes all regions equally deep. The sign survives only if depth is counted by who went\nthere, which the count of § 4.2 gives.\n\n---\n\n## 4. The controls for a deeper step\n\nControls 1 to 5 of § 1.1 carry over as they are. What follows is what depth adds to them.\n\n### 4.1 Look it up before building under it\n\n**At every level, a made sentence is looked for before anything is written under it**: on the map,\namong the claims of the two texts, and in outside sources on both sides. If a text states it, the\nmachine points at that claim, and what lies under it is ordinary: an author's claim with its own\nlinks and questions. The count of made levels starts again there.\n\nThis is the brake that works against § 3.1 directly. A chain of five made sentences is five\nmachine steps. A chain that reaches a stated sentence at its second level is two.\n\n**The worst error in this design is a made sentence marked as found.** It lends a text's standing\nto the machine's words, and it resets the count falsely. So *found* is a small task of its own,\nreported by two rates as the open row on checks asks (Y12): of the sentences a text does state,\nhow many it found; of those no text states, how many it wrongly pointed.\n\n**Both can be tested today with no person.**\n\n| Answer key | Size, counted 2026-10-10 | What it tests |\n|---|---|---|\n| Claims that carry their own quoted passage. The passage is the author's words and the claim is a rewording of it: the middle overlap of content words between the two is 0.26 | 1,789 claims in 34 texts | hand the pointer a passage and the text's claims: does it point at the claim made from it? Take that claim away: does it say *none*? |\n| Chains an author wrote: a first claim supports a middle claim, which supports a conclusion, all in one text | 200 chains of two in 20 texts (70, 34 and 29 of them in three texts); 85 chains of three in 10 texts | hide the middle and hand over the two ends: does the pass find the author's own middle before it makes one? |\n\nBoth counts: [chains_authors_wrote.py](quality-control-for-deeper-steps/chains_authors_wrote.py).\nEach is a test of recovery. Neither says a found sentence is true.\n\n**What is known about pointing so far.** On 7 October, 119 pointers all named a claim that exists,\nwhich code checked. Whether each named claim says what the sentence says was read by Claude only;\na first try had counted a claim about one case as stating a general rule, and one denial was\nmisread. The quote check that exists confirms that words are in a source. It does not confirm that\na claim says what they say (ledger Y6).\n\n**A fetched passage is checked by code to stand in the fetched text.** The design of 25 August\nnames fetched research as where invented references arrive.\n\n### 4.2 Every made sentence carries its count and its record\n\n| It carries | Why |\n|---|---|\n| **How many made sentences it stands on**, counted from the nearest sentence that a text states, a located passage, or something a person wrote or took up. A link the machine drew counts as one | this is the dial of his ruling of 29 September, per sentence. It is also what keeps depth a sign of care (§ 3.6) |\n| The sentence it was written under | so that undoing a sentence can mark everything below it |\n| The number of its recorded judgment | *based on what* has an exact answer: the texts that record lists |\n| The writer's version and the model that answered | a rate belongs to one version of one task on one model |\n| The kind of ground it was written as (§ 4.4) | the check that fits depends on it |\n| Each check's verdict, and which checks could not be run | a sentence no check fits says so |\n\nThe count is stored and shown wherever the sentence is shown.\n\n**Where it would be kept: the graph has no place under a link today** (read from the live graph\n2026-10-10, on the founder's question whether the rules are stored as claims). They are not in\nthe graph: every sentence a machine has written under a link in the trials is in the trial files,\nin the review batch (§ 5.4) and in the record of machine judgments, which has kept each model call\nsince 6 October with the texts it was shown. The graph has five kinds of node (6,417 claims, 1,225 standing questions,\n100 concepts, 35 sources, 3 readers) and fifteen kinds of link, nine of them between claims. A\nclaim is the only node that carries content. A link is an arrow with fields on it (its pattern, a\nstrength, a reason), and in this kind of database nothing can be attached to an arrow.\n\n⚠ **Corrected the same evening** (`claim-or-system.md` § 1). This paragraph said the standing\nquestion is the one thing kept under a link, and offered three places a made sentence could go.\nRead again, four things sit under a link today. Two are fields on the arrow that nobody can point\nat: the linking step's own reason for the link (on 1,481 of 1,636 support and attack links between\nclaims from texts or people) and a machine's reading of relevance (190 links). One is the standing\nquestion, a node holding both claims' ids. And one is **a claim about the link, whose id the arrow\nholds**: the relevance claim, built on 28 September with the two-factor ruling, attackable like any\nclaim, and never yet made on the live graph. So the three places were one place seen three ways.\nThe proposal that follows from it, that a machine's sentence is a claim nobody asserts, kept inert\nunder the link's relevance claim, is ledger AG12. **Ruled that evening** (founder: *\"The moment\nit's written. Go ahead and make it happen.\"*) **and built**: what the graph holds under a link now,\nand what was left unbuilt, is `claim-or-system.md` § 4a and the drawing in\n`docs/architecture-reference.md` § Graph Database.\n\n### 4.3 One level to a call, and the writer sees the sentence alone\n\n- **No chain is written in one go.** In the one outside measurement the corpus has, whole chains\n  came out right 79 times in 100 at one step, 18 at two and almost never at three (an older model,\n  strict scoring, a helper's reading).\n- **The writer of a level is shown the sentence it opens and nothing of how that sentence was\n  written or checked.** This is the remedy the outside finding of § 3.1 points at. The record\n  lists each call's inputs, so it can be checked by query.\n- **A checker is shown the passages and the two end claims, never the machine's steps above the\n  one it checks.**\n\n### 4.4 Two questions before anything is written, and the answer decides the form\n\n**Asked as one question, *why does this hold* got one kind of answer.** In the trial, the choice\nincluded *it has been seen to hold in case after case* and *a body with standing decided it*. In\n38 askings these were named once. The model read *why* as asking what makes the rule so. How the\nrule is known is a second question, and the corpus already keeps the two apart: the evidence note\nproposes evidence of mechanism as its own stream beside evidence from cases (ledger Q8), and the\nphilosophy note has Bolzano's two relations, one of which never runs from a sign to its cause.\n\n*Added 2026-10-10, while the day's walkthrough was written.* The distinction has an older form\nthat the corpus had not named. Aristotle separates a proof of the fact from a proof of the\nreasoned fact, with the planets as his example: that they do not twinkle shows that they are\nnear, and their being near is why they do not twinkle (*Posterior Analytics* I.13, read as a\nsearch summary of the chapter and not in full). The first sentence runs the way a thing is known\nand the second the way it comes about. The founder's squirrel of 2011 has the same two directions:\nthe smell is how the death is known, and the death is what makes the smell.\n\n**So two closed questions are asked of a made sentence, apart, each behind a gate that asks first\nwhether any listed answer fits: what makes it so, and how is it known.**\n\n| The answer | What the machine does | What checks it | State |\n|---|---|---|---|\n| **A wider rule** | files it as a wider wording in the rule's own record (ledger AG4). It is not a level down (§ 5) | the first-level reader, which needs same-subject alternatives it does not have here | tried on 19 |\n| **How it comes about** | writes the chain, one stage to a sentence, as questions; looks every stage up, on both sides | finding the stages. § 5 has what leaving a stage out showed | tried on 10 |\n| **Seen in cases** | finds cases in texts, or takes them from people. **It never makes one.** It counts them by code and gives no figure | the case is located; the count is exact | counted before: 84 links file a case under a general claim |\n| **A source says so, or a body decided it** | looks the source up | code: the passage is in the source | unbuilt |\n| **What its words mean** | hands it to the concept layer: whose sense, stated where | none here | unbuilt |\n| **What matters** | hands it to the weighing questions. On a person's own body, spirit or sacred commitments: nothing uninvited | none here | the questions ship |\n| **None fits** | says so, and leaves the question with a move beside it | | the gate never fired in the trial: § 5 |\n\n**A made case counts for nothing.** The hand that wrote a rule, imagined the case for it and\nchecked the step is one witness asked three times (ledger AD27, and § 5.2 of the note this one\nfollows).\n\n**The descent asks against as well as for.** Under every made sentence the question *what would\nbreak this* stands beside *why does it hold*, as the ruling of 25 August asks of research. A case\nagainst counts when it is found in a text, brought by a person, or raised by a model of another\nfamily.\n\n### 4.5 The chain is checked as a chain\n\nOnce a made sentence stands on a made sentence:\n\n| Check | By | Catches |\n|---|---|---|\n| No sentence says a sentence above it, or either claim of the link, over again | code, by shared words; then the sameness check | a circle |\n| Each hop's first half is written out: does the sentence above support this one | the first-level form and reader, for hops of that shape | a middle that is only another argument for the conclusion (§ 3.4) |\n| Each hop records which way it runs | the pattern classifier, at the gate | support manufactured by mixing directions |\n| The descent is written twice, apart. Where the two agree the chain is marked steady. Where they differ both are kept, and nothing is written under the difference until it is read | the writer twice, then the sameness check | a chain that changes when it is made again |\n| Nothing in the tree is denied by another sentence in it, or by a claim of either text | pointing, with code checking the claim exists | a chain that contradicts itself. A denial by a text is offered as the crux |\n| From the lowest sentences and the first claim, a reader not told the link reaches the second claim | the first-level reader | drift, for chains of the first shape |\n\nAt the first level the two writings of a rule were the same, or one contained the other, on 23 of\n27 links. Nothing is measured at a second level.\n\n### 4.6 Nothing else stands on a made sentence\n\nUntil a text states it or a person takes it up:\n\n- the linking step does not draw links to or from it;\n- the sameness check does not merge it with a claim, and does not offer it as a merge;\n- the strength code does not read it;\n- no summary, map default or agent surface serves it as a claim.\n\nOnly the descent and its checks read it. **One test reads the code for a reader that forgets**, as\none already does for replaced part links. When a text states the sentence or a person takes it up,\nit becomes an ordinary claim, with that person or that text as its source, and the count stops.\n\nThis is what keeps § 2 true, and it is what lets his ruling stand beside the splitter's rule.\n\n### 4.7 Brakes that are not the writer's\n\n**The writer does not stop.** It wrote a rule for 29 and 28 of 30 links on 7 October. In § 5 it\nwrote a wider sentence for 19 of 19 rules. Its own *no* fires only where the form asks for\nsomething concrete.\n\n**The stop rule of 25 August is a weak brake on fresh material.** The rule is: stop where nothing\nmore could change what a reader sees. As built (`deliberus/question_ordering.py`,\n`deliberus/graph/qbaf.py`), a question is worth asking when a decisive answer would move its claim\nacross a display band. A claim nobody has examined sits in the grey band, so any answer that gives\nit a first reading counts, however small the movement.\n\nMeasured on the live graph, 2026-10-10\n([the script](quality-control-for-deeper-steps/what_the_stop_rule_says.py)):\n\n| Open questions on a link into a claim | Priced worth asking | Priced not worth asking |\n|---|---|---|\n| the claim not examined yet | 595 | 603 |\n| the claim has a reading | 0 | 25 |\n\nSo on fresh material it lets half through. The half it stops, it stops because the claim at the\nother end of the link has no standing yet, so the answer would move nothing: a reason that has\nnothing to do with how likely the machine is to be wrong. And it reads standing questions, which\nexist only on links that carry a pattern: 930 of the 1,569 links between claims carry none, the 822\nthe machine drew among them, so under those it has nothing to read. The earlier reader's finding\nstands, sharper: the August design has a stop by value and no stop by error.\n\n⚠ *This paragraph first said that on fresh material the rule always says continue. That was read\noff the code. Run, it stops about half.*\n\n**So a level ends in one of four ways, none of them the writer's choice:**\n\n1. a check fails, and the branch ends as *cannot be written*, with its question;\n2. a ceiling of calls for one link and for one run is reached. The linking step has such a ceiling\n   since 2026-10-07 and reports *held*;\n3. the order of work puts other links first: links somebody disputes, rules many links use, links\n   where an answer would move most;\n4. once it is measured, the limit of § 4.8.\n\n**Undoing reaches down.** Undoing or replacing a made sentence marks every sentence written under\nit as owed a new reading. The splitter's undo works on one claim and would have to follow the\nrecorded parent.\n\n### 4.8 How often it is wrong, measured by level\n\n**The limit on made depth is a measurement, and the measuring is the quality control he asked\nfor.** In order:\n\n1. **Now, with no person**: the two tests of § 4.1, and the five tests that need no answer key\n   (ledger AF2).\n2. **A person reads first** (his principle of 6 October). A sitting for steps: one step beside its\n   link and its two passages; the first thing wrong; four answers (enough; something is missing; it\n   says a claim over again; this is not what the link rests on) and a note. The items are drawn at\n   random across levels and kinds, since people left to choose will read what they dispute, and\n   that is the rarest act in the system. One item in five is shown without the machine's own\n   verdicts, to measure how far those sway the reader.\n3. **The faults are grouped and the task is rewritten.** An error rate belongs to the task as\n   written before it belongs to the model.\n4. **A rate is proven** per version, per kind of ground and per level: blind, random, about 60\n   clean items for under 5 in 100. A level-two rate is measured on steps built from the sentences\n   actually stored at level one.\n5. **Then the limit**: the largest count of made levels whose summed rates stay under his\n   tolerance. It differs by kind.\n\n**Until step 5, the pass writes as deep as its checks pass**, within the brakes of § 4.7, every\nsentence carrying its count, unseen by default.\n\n**After that, its own mistakes are the probe.** A made sentence that a person disputes, or that a\ntext is later found to deny, is logged as a miss at its level and kind. A level whose rate rises\nabove the tolerance stops being written. He dropped planted errors from the live graph on\n2026-08-31 (*\"I don't think the seeded-probe thing sounds useful and elegant and worthwhile, do\nyou?\"*), so planted errors stay inside sittings.\n\n**Two cautions the corpus already holds.** A count of objections means nothing if the page that\nshows the steps is not being opened: the splitter's correction record stood empty for a week\nbecause its panel did not render, not because nobody objected. And an audit needs a count that\ndoes not depend on who happened to complain: eligible, sampled, examined.\n\n**His words on where this ends**, 29 September: *\"it's a fine balance between challenging the\nstatus quo and SoTA beliefs about how far and deep this can go, vs accepting where the edge of\npossibility lies and empowering humans to review and contribute beyond it (or to check the\nmachines' mistakes without being overwhelmed).\"* The curve of faults by level is how the edge is\nfound. Beyond it the machine leaves the question, with a move beside it.\n\n### 4.9 What a reader is shown\n\nUnchanged from the note this one follows: hidden by default, shown on contest, each sentence a\nquestion with a move beside it, with its count. Three things the readers found that bear on it,\nnone settled here (§ 8): a machine's mechanism does not do for its reader what working one out\ndoes for the person who works it out; a ratio shown as a figure can read as a grade; and a person\nasked for their own reason should be asked before being shown the machine's.\n\n---\n\n## 5. A first look at the second level\n\n**What was tried.** The nineteen rules kept after pruning on 7 October. Each is a sentence a\nmachine wrote under a link the linking step drew. One model throughout (`claude-sonnet-5` on the\nkeeper's plan, strict, the answering model recorded on every call). Nothing was written to the\ngraph. Predictions were written to a file before the first call\n([predictions](quality-control-for-deeper-steps/second_level_predictions.txt);\n[the script](quality-control-for-deeper-steps/second_level_trial.py);\n[the answers](quality-control-for-deeper-steps/second_level_trial.json);\n[what the scoring printed](quality-control-for-deeper-steps/what-the-runs-printed.txt)).\n\n1. **What kind of answer is owed.** One closed question, six kinds behind a gate, asked twice with\n   the list in opposite orders.\n2. **The wider sentence**, for every rule: the sentence the rule would be a case of, and the tie\n   between them, with leave to write none. Checked as a first-level step was: by code, and by a\n   reader not told what the sentences are meant to establish, who picks among four rules or none,\n   in both orders, then with each part alone.\n3. **The chain**, for the ten rules that either asking read as resting on how something works: two\n   to five stages, with leave to write none. The same reader got the chain in order, reversed, and\n   with each stage left out.\n\n### 5.1 Against the predictions\n\n| | Predicted | Came back | |\n|---|---|---|---|\n| P1 | the two askings agree on at least 13 of 19 | 14 | held |\n| P2 | *a wider rule* on at most 3 | 4 | **failed** |\n| | *what its words mean* the commonest, at least 5 | 4; the commonest is *how it works*, 6 | **failed** |\n| | *how it works* on 2 to 5 | 6 | **failed** |\n| | *what matters* on 2 to 5 | 0 | **failed** |\n| | *seen in cases* on 1 to 4 | 0 | **failed** |\n| | *decided* on none | 0 | held |\n| | *none fits* on at most 3 | 0 of 38 askings | held, and see below |\n| P3 | a wider sentence is written for at least three quarters of the rules that neither asking read as resting on one | 12 of 12 | held |\n| P4 | the reader reaches the rule from the wider sentence and the tie on at least 80%, the same in both orders on at least 80% | 16 and 18 of 19; the same answer on 17 | held |\n| P5 | each part alone still reaches the rule on at least a third | the tie alone 9 of 19; the wider sentence alone 13 of 19 | held |\n| P6 | in at least half, half the rule's own words come back | 4 of 19; the middle share is 0.18 | **failed** |\n| P7 | a chain is written for every rule asked | 8 of 10; two declined, each with a ground | **failed** |\n| | the middle chain has 3 to 5 stages | 4 or 5 stages in every chain | held |\n| | the reader reaches the rule from the whole chain on at least 70%, the same in both orders | 8 of 8 in order, 8 of 8 reversed | held |\n| | with one stage left out it still reaches the rule on at least half | 36 of 36 | held, and see below |\n\nCounting each part, sixteen things were predicted: nine held and seven failed.\n\n### 5.2 What it says\n\n**The gate never fired, and that is an alarm.** *None fits* was available at every asking and was\nchosen 0 times in 38. Four of the nineteen rules say what ought to be done or what is just (rules\n6, 20, 25 and 28). *What matters* was named for none of them in either asking, and three were read\nboth times as resting on *a wider rule*. The list mixed two things, the form of an answer and its\nsubject, and the model chose by form. The fault is in the list. By the principle on guards, a zero\nhere is reported and not passed.\n\n**The list's order moved a quarter of the answers.** Five of nineteen changed, four of them toward\n*how it works* in the asking that listed it ahead of *a wider rule*. The plain case is the rule\nthat *producers avoid paying more for an input than the revenue it earns them*. In the one-rule\ntest run it was read as one case of profit-seeking under one order and as a chain from overpaying\nto loss to avoiding it under the other; in the main run both askings read it as a chain. Both\nreadings are sound. The rule has both kinds of ground, and one question with one answer hides\nthat. This is a second reason for asking the two questions of § 4.4 apart.\n\n**That a deeper step can be written says nothing, here as at the first level.** A wider sentence\nwas written for all nineteen. Twelve of those were rules that the same model had twice read as\nresting on something else.\n\n**A wider sentence is the rule made wider, and no easier to check** (Claude's reading of all\nnineteen; nobody else has read them). It is seldom the rule's words over again: the middle share\nof words that return is 0.18. It is a larger claim of the same kind. For *a significant,\neconomy-wide increase in productivity generally gives rise to broad economic abundance* the writer\ngave the same sentence about total output. For *acts that differ only in how they are classed do\nnot differ in whether they advance justice* it gave *…do not differ in any of their moral\nproperties*, which claims more and is harder to believe. A reader who doubts the rule doubts the\nwider sentence at least as much. So going wider does not carry a question toward anything that\ncan be looked at. It is what the earlier note already files in a rule's own record, as its wider\nwordings (ledger AG4).\n\n**The first-level check leaks here.** Of the 16 steps where the reader reached the rule in both\norders, both parts were needed in 3. The wider sentence alone was enough in 5, the tie alone in 2,\neither in 6. At the first level the reader chose among four claims of one text, all on one\nsubject. Here the other three choices were rules from other links, on other subjects, and a reader\ncan pick by subject. The tie names what the rule is about, which would explain its being enough\nalone on 9.\nThis is a limit of the trial's design, and it is also the finding: **under a made rule there is no\ntext to take same-subject alternatives from.** A check of a second-level step needs alternatives\nmade for it (the rule turned round, the rule made stronger, the rule denied), and those are\nfurther machine sentences.\n\n**Asked for a chain, the writer said no twice, and said why.** For *social attitudes that bear on\ncooperation are typically produced by a specific underlying mechanism*, it answered that the rule\nonly asserts that some unnamed mechanism exists, so nobody could say the route. For *the workspace\na system draws on is the same workspace in which it organizes its response*, it answered that the\nrule claims two things are one thing, which is not one thing bringing about another. Both rules\nhad been read as resting on *how it works* in both askings. So the first question and the writer\ndisagree on 2 of 10, and the writer's refusal is the better grounded: the form asks for a thing\nthat brings about another thing, and where the rule names none the form cannot be filled. A\nwriter's *no* fires when the form asks for something concrete. Asked for a wider sentence, which\nany sentence has, it never fired.\n\n**Leaving a stage out showed nothing.** The reader reached the rule from all eight chains in both\norders, and from all 36 versions with one stage missing. No stage of any chain was needed by this\nreader. With three alternatives on other subjects, one or two stages are enough to tell which rule\nis meant. So this check cannot tell a complete chain from one with a hole. Nothing built can. For\na chain the check is finding each stage in a source, and a person's reading.\n\n**What a chain is good for shows in one example.** For *a significant, economy-wide increase in\nproductivity generally gives rise to broad economic abundance*, the wider sentence was the rule\nagain with *total output* in place of *productivity*. The chain was: more output from the same\nlabour and capital; a lower real cost for each unit; competition among producers and workers'\nbargaining force firms to pass the lower cost on as lower prices and higher wages; an hour of work\nbuys more; and where that reaches households generally and does not stay with a few, it is broad\nabundance. The third and fifth stages carry the two conditions the rule depends on: that the gain\nis passed on, and that it is widely spread. Each is a claim a source can be found for or against,\nand each says where the rule stops holding. That is the earlier note's point that a mechanism sets\na rule's reach, seen on a made example.\n\n**Not every chain written was a chain of what brings about what** (Claude's reading). Five of the\neight are (rules 6, 11, 13, 16, 29). Three are an argument or a calculation set out in stages:\nrule 30 is an inference by ruling one explanation out, rule 12 is arithmetic over rates, and rule 8\nis close to what its own words mean. The form was stretched to fit them. A person's reading would\nsay whether that sorting is right.\n\n### 5.3 What the trial cannot say\n\n- Nineteen rules, one model, one evening. No person has read a wider sentence or a chain.\n- The rules sit under links the linking step drew, seven of which a hand reading had called no\n  reason at all. Rules under an author's own steps may behave differently.\n- The kinds of ground are a list written that afternoon. It broke on its first use, as lists do.\n- The run made 203 model calls on the keeper's plan, a one-rule test run included: about $12 at\n  list prices. **No money was paid**; the figure measures plan allowance.\n\n### 5.4 The page where a person reads them (built 2026-10-10)\n\n**What it is for.** Step 2 of § 4.8: a person reads what the machine wrote, before anyone proves\nhow often it is wrong, and writes down the first thing wrong in each case. Asked whether to build\nthat page or to run two more tests that need nobody, the founder answered: *\"Build the page\"*.\nThe step review is that page: `/step-review/first-reading`, its API at `/step-reviews/…`\n(`deliberus/step_review.py`, `web/src/routes/step-review/`). Live since 19:35 that evening.\n\n**What it holds.** Every sentence a machine has written under a link so far, which is two\nevenings' trials: under the thirty links of 7 October, 29 written steps and one refusal; under\nnineteen rules of those steps, 19 wider sentences, 8 chains and 2 refusals. 59 items. When the page was built nothing a\nmachine wrote under a link was in the graph, so the items\ncome from a file made out of the trials' stored answers (`scripts/build_step_review_batch.py`,\nwhich refuses if a claim's text on the live graph differs from the trial's). Since the ruling of\nthat evening (ledger AG12) the same file's sentences are also made claims, word for word, by\n`scripts/write_machine_sentences.py`. The page still reads the file. Reading a batch from the graph\nis not built.\n\n**What a reader is shown, one item at a time.** The two claims of the link, each with the passage\nit was taken from and a way to open it. Under them, what was asked and the sentences the machine\nwrote, each marked with the kind the writer said it is. A second-level item also shows the\nfirst-level step with the rule in question marked, so a rule is never read alone. Then the\nmachine's own checks, in words: what they found, which sentences the check needed, and for the\nsecond level the caution of § 5.2 that reaching a rule shows less there.\n\n**The answers.** The four of ledger AD20 (enough; something is missing; it says a claim over\nagain; this is not what the link rests on), and two additions that are Claude's and unruled:\n\n- *There is nothing to write: the first claim is no reason for the second* (and under a rule:\n  *the rule itself does not hold*). A hand reading called seven of the thirty links no reason at\n  all, and the linking step drew them. Filed under *not what the link rests on*, that fault would\n  count against the writer of the step when it belongs to the step above it, and an error rate\n  belongs to the task that made the error.\n- A second set of answers where the writer declined, since the question there is whether\n  declining was right.\n\n**The note is the product.** Its box is open, and it asks for the first thing wrong in the\nreader's words. A reader can also point at one sentence as the one that is wrong: a tap, for the\ncase where something is off and the words for it are not there yet.\n\n**One item in five comes without the machine's checks**, withheld by the server as in the two\nolder reviews, to measure how far the checks sway the reader.\n\n**The order differs from § 4.8, on purpose.** That section said items are drawn at random across\nlevels and kinds. Here the *links* are drawn, and a link's items come together: the step, then\nwhat is under its rule. The page still chooses and not the reader, which was the point of the\ndraw. A reader meets the items in the order a map would unfold them, and reads two claims once\nfor up to three items. The cost: answers about one link are not independent of each other.\n\n**What is stored and what is public.** One row per reader and item in `step_answers`, carrying\nwhat the reader was shown as well as what they said. `GET /step-reviews/{name}/results` gives\ncounts only: by level, by form, by what the checks had found, and the reader against the checks,\nshown and blind. Notes are never served. They are read with `scripts/read_review_notes.py`.\n\n**What it cannot say.** This is a first reading. It finds kinds of fault and proves no rate: 59\nitems, half of them from two texts on one debate, every second-level item under a rule that had\nalready passed the first level's checks, and one reader until a second sits the same name. To try\nthe page without writing to the live answers, `scripts/mock_review_api.py` serves it locally.\n\n**How it was checked before it went live.** The tests stand in for the database, so the new table\nand its two functions were also run against a throwaway Postgres 16 on the keeper's machine, through\nthe real routes: made at start, an answer saved, replaced and read back, the results naming nobody\nby address and carrying no note. In the browser at desktop and phone width, with saves made to\nfail: an answer refused by the server is said to be lost and comes up again before the end.\n\n---\n\n## 6. What this changes in earlier rows and notes\n\n| Where | As written | Now |\n|---|---|---|\n| [the-chain-inside-a-link.md](the-chain-inside-a-link.md) § 4.3, § 5.1; ledger AG2, AG5 | below its first level the machine finds and asks; with no measured rate it makes one level | withdrawn by AG6. Found first at every level; made where nothing is found, under the controls of § 4 |\n| Ledger T1, second half | *with no measured rate a person engages first* (never ruled) | for this pass he ruled the other way, as he did for the splitter: writing marked first, the sitting after. The ruled half stands: measurement caps how deep machine work builds on machine work |\n| The same note, § 4.3 | *a found sentence is one machine step from its text* | it stands as far from its text as the reading that took the claim out and the pointing that found it |\n| The same note, § 4.6 | *today the arithmetic pays for the leap* | on a link a text or a person drew. A link the machine drew is presumed to bear nothing, so nothing written under it can lower it |\n| Ledger AD19, AD29; [writing-out-the-unwritten.md](writing-out-the-unwritten.md) § 9.7 | depth by taking a rule in its turn, capped by measured error | *no chain in one go* stands as a control (§ 4.3). The cap is a measurement to be made (§ 4.8) |\n| Ledger AG4 | a rule is kept as every wording proposed for it | gains a source of wider wordings: what the writer gives when asked what the rule is a case of (§ 5) |\n| Ledger Y1 | 5 to 10 in 100 for a marked item no later machine work stands on | does not cover a made sentence with one under it; the bar is on the chain (§ 2) |\n\n**The rule of Claude's that did not fire.** His sentence of 7 October asks for all the unwritten\nstructure to be written out *\"to the extent that process can be trusted\"*, depending on\n*\"microtasks that we can rely on deeply\"*. His position of 29 September is that an error rate is a\ncost to be driven down by better tasks, and that the aim is to find where the edge lies. Both were\non the page of the note that proposed *find and ask only*. The pick should have been predicted\nfrom them.\n\n---\n\n## 7. Proposals (all unruled; ledger AG7 to AG11)\n\n| # | Proposal | § |\n|---|---|---|\n| AG7 | **A deeper step is the machine's question under a link, and nothing but the descent reads it.** It is never part of anybody's claim, it weighs and caps nothing, and until a text states it or a person takes it up no other pass links it, merges it, weighs it or serves it. A test reads the code for a reader that forgets. | 2, 4.6 |\n| AG8 | **The machine looks a made sentence up before it builds under it, and every made sentence carries how many made sentences it stands on.** *Found* is a small task with two rates of its own, tested now on 1,789 claims that carry their own passage and on 200 chains authors wrote. | 4.1, 4.2 |\n| AG9 | **Two closed questions are asked of a sentence before anything is written under it: what makes it so, and how is it known. The answer decides the form and the check.** A wider rule goes into the rule's own record. Cases are found and never made. One level to a call, and the writer sees the sentence alone. | 4.3, 4.4, 5 |\n| AG10 | **A chain is checked as a chain, and its brakes are not the writer's.** No circle; each hop's first half; each hop's direction; written twice; nothing denied inside. A level ends on a failed check, a ceiling of calls, or the order of work, and undoing a sentence marks what is under it. | 4.5, 4.7 |\n| AG11 | **The limit on made depth is measured by level, and until it is the pass writes as deep as its checks pass, counted and unseen by default.** Tests with known answers; a person's first reading, random across levels; a rate per version, kind and level; then the limit. | 4.8 |\n\n**Read by the outside reader** (a model with none of this session's context, comparing each row\nwith everything he has settled; `scripts/outside_check.py`). It found no reversal on any of the\nfive rows, nor on the two earlier rows reworded for his ruling (AG2, AG5). It found one narrowing:\non AG7 it reads *the sameness check does not merge it* as narrowing the ruled order that sameness\nis decided before any relation is written. Claude's answer is on the row: that order governs pairs\nof claims, a made sentence is no claim, and the descent uses the sameness check to look a made\nsentence up and to compare two writings. Which reading governs is his. Each row carries both\nlines, the author's and the reader's.\n\n---\n\n## 8. What this does not settle, and what could go wrong\n\n- **No person has read a written-out step, at any level.** By his principle a person reads before\n  anyone proves. The thirty first-level steps and this trial's nineteen are waiting, and no page\n  shows them.\n- **People's reading is itself fallible, in ways the corpus has measured elsewhere**: in one\n  study four in five people who detected a machine's error followed it anyway; raters' agreement\n  on whether a step is sufficient is low in the outside figures; and a small group of like-minded\n  readers measures its own agreement. Hence blind items, a second sitter, and the random draw.\n- **Hidden by default, against showing depth.** The ruling of 25 August keeps the machine's\n  descent off the default map. The note on curiosity says the mitigation for depth is not to hide\n  it. And the principles page says that knowing one could check raises trust even when one does\n  not, which for a machine's unread steps is a risk and not a comfort. Not settled here.\n- **A share shown as a figure.** The earlier note proposes publishing the share of links with\n  nothing under them (AG1). The principles page declined a ratio once: *\"One of six used\" is a\n  score wearing a disclosure's clothes*. A corpus figure on a statistics page is not a grade on\n  one answer, and the tension is his to weigh.\n- **Things left unsaid on purpose.** The carve for a person's own body, spirit and sacred\n  commitments rests on a detector that fires on two claims in the corpus. And a text several\n  parties negotiated can hold an ambiguity that does work for them: the corpus has no notion of\n  that, and *a text's own steps are answered eagerly* would write it out. Not settled here.\n- **A made question costs a reader something too.** In a 2025 shared task the best systems' critical\n  questions were useful a little over half the time, and model classifiers rated three quarters or more\n  of their own candidates useful where under half were\n  ([derivability-of-missing-considerations.md](derivability-of-missing-considerations.md)). A descent\n  that asks against as well as for will offer questions of that quality until a person's reading\n  says which kinds to stop offering.\n- **The same thought, earlier in the corpus.** That the deeper the descent goes the more machine\n  output feeds machine input was named in August from a paper on learned world models\n  ([grounded-critics-and-quarantined-imagination.md](grounded-critics-and-quarantined-imagination.md)),\n  and that many plausible steps from one generator are one sample repeated is the first class of the\n  error taxonomy ([the-residual-error-taxonomy.md](the-residual-error-taxonomy.md)).\n- **One family of models all the way down.** A second family for the reader was not used, since it\n  would spend his money. Where the corpus has measured it, a second family helps and does not make\n  the checks independent.\n- **Cost.** The first level cost about 17 calls a link on 7 October. A level under each kept\n  sentence costs about that again. The descent is affordable only in the order of work, and the\n  ceiling of § 4.7 is what holds it there.\n- **The strength code has fixed depths** (8 levels of parts for a badge; 6 levels and 200 claims\n  for the sensitivity reading). Made sentences are outside it. A chain that people take up level\n  by level could reach them.\n\n---\n\n## 9. Found beside the question\n\nTwo things in other places that the reading showed to be out of date. Each is kept in the corpus\ngraph and listed in `TODO.md`; the first is written into its note.\n\n- **The guard against thinly spread error was called unbuilt in July and has had a first form\n  since 22 August** ([threads-synthesis-and-gaps.md](threads-synthesis-and-gaps.md) C3;\n  [slicing-stability-first-run.md](slicing-stability-first-run.md)). That form refiles the same\n  parts four ways by code. It does not make a decomposition a second time, which is the check of\n  § 4.5.\n- **The splitter's rebuilding of a claim from its parts was described as blind or unanchored in\n  five places** (§ 1.3, first hole): the module's header, two always-loaded lines,\n  [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) § 3 and\n  [a-cut-through-a-concept-is-an-argument.md](a-cut-through-a-concept-is-an-argument.md). Each now\n  says what the call is shown.\n\nA third was reported and did not hold: a reader found a quoted passage in the April note on the\ntruth graph saying that strength converges toward probability, which the ledger superseded. That\nnote already carries the mark, a hundred lines above the quotation.\n\n---\n\n## 10. The corpus graph: what it held, and what was added\n\n**Asked first.** Before the readers were sent, the graph was asked about 42 rows and notes. Its\nanswers named 67 notes.\n\n**The forward test its own note asks for** ([corpus-graph/README.md](corpus-graph/README.md) § 8):\nthe three readers' reports cite 77 notes. The graph had named 33 of them. The other 44 it had not,\namong them the notes on what people's checking can be relied on for, the public page on depth, and\nthe red-team notes. The count is crude: it includes notes a reader opened and found nothing in.\n\n**Kept.** Twenty-three connections, each with who found it, why it matters, and the passage it rests\non at both ends (`corpus-graph/found.jsonl`, the lines whose `where` names a report of this note). Twenty-one\nrun from this note to an older one, one from the note this one follows to the principles page, and one\njoins two older notes. By kind: seven narrow a passage, five supply a piece or a number or a check, four\nare tensions left for the founder, two are the same idea earlier, and one each is a later correction, a\ncontradiction, a reconciliation, an earlier answer and a test for a waiting prediction. Three were also\nwritten into the older note the same day, and are marked so: the first form of the stability check, the\nstop rule in the August design, and the splitter's check. A session that opens any of the other older\nnotes is now told of the connection.\n\n**The readings are recorded**, so that `around` says a reader has been through the note this one\nfollows, and for what.\n\n**What the three readers cost**: about 2.05 million tokens of a mid-size model on the keeper's\nplan, 439 tool calls, 17 to 26 minutes each. He asked the same evening, in another session,\nwhether sending readers is good practice now that the graph exists. That session's answer is in\nthe graph's note (§ 12): readers earn their cost on notes nobody has read for, when what they find\nis kept. These three were sent before the graph could say which notes had been read for.\n\n---\n\n## 11. What was read, and how far\n\n**The three readers' reports, word for word**, with their briefs:\n\n- [how the small tasks are checked](quality-control-for-deeper-steps/corpus-bearings-how-the-small-tasks-are-checked-as-a-reader-returned-it.md)\n- [depth, chains, error and volume](quality-control-for-deeper-steps/corpus-bearings-depth-chains-and-error-as-a-reader-returned-it.md)\n- [people, showing, and what checks cannot see](quality-control-for-deeper-steps/corpus-bearings-people-and-what-checks-cannot-see-as-a-reader-returned-it.md)\n- [the three briefs](quality-control-for-deeper-steps/the-three-briefs.md)\n\n**Which of their findings were opened afterwards, on both sides, and used.** In the code: the\nsplitter's check and its prompt; the sameness check's order control and how linking calls it; which\nlinks the reader of bearing selects; the list of presumed writers; how the stop rule prices a\nquestion; the fixed depths. In the notes: the passage on screening off; the prediction on drift;\nPearl's abstract as a helper recorded it; the two lines of the page on depth; the splitter's rule\non found and added; the end-state of 31 August and the dropping of planted errors; the passage on\ndetecting and deferring; the empty correction record; the count that does not depend on who\ncomplains; the first form of the stability check; the superseded line on probability. Findings\nthat were not opened are not used above.\n\n**Read directly by Claude, this session.** [proving-a-pass-reliable.md](proving-a-pass-reliable.md)\n§ 0 to § 7, § 9; [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) § 0 to § 2,\n§ 3e, § 4, § 5; [the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md)\n§ 1 to § 3b; [earning-the-right-to-write.md](earning-the-right-to-write.md) § 1, § 3, § 3a, § 4,\n§ 4a, § 5, § 6; [writing-out-the-unwritten.md](writing-out-the-unwritten.md) § 6 to § 11;\n[philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 5;\n[what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 6;\n[the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md) § 4; the\nledger rows T1 to T4, Y1 to Y14, R8, R12, Z1b, AD20 to AD29, AF2, AF4; the helper's outside report\nof 7 October on dependable small tasks, in the part on chains.\n\n**In the code.** `deliberus/judgments.py` (header); `deliberus/relevance.py`;\n`deliberus/relevance_judge.py` (which links it selects); `deliberus/question_ordering.py`;\n`deliberus/graph/qbaf.py` (the display bands); `deliberus/extraction/decompose_claim.py` (the check\nand its prompt); `deliberus/keeper.py` and `deliberus/extraction/auto_connect.py` (the order\ncontrol and the *same* branch); `deliberus/parts_review.py` and `deliberus/sameness.py` (the\nsittings' blind and planted shares); the trial script of 7 October.\n\n**Counted on the live stores, 2026-10-10, no model.** Chains authors wrote, and claims that carry\ntheir own passage (2,050 extracted claims, 1,789 with a quoted passage of six words or more):\n[chains_authors_wrote.py](quality-control-for-deeper-steps/chains_authors_wrote.py). Answers in\nthe two sittings, notes, and machine-cut parts a person has touched:\n[what_people_have_read.py](quality-control-for-deeper-steps/what_people_have_read.py). Everything\nboth printed: [what-the-runs-printed.txt](quality-control-for-deeper-steps/what-the-runs-printed.txt).\n\n**Outside sources.** None was read for this note. The outside figures above are from two helpers'\nreports of 7 October, each of which gives its own read depth, and are marked where used. Verify\nbefore citing outward.\n\n## See also\n\n[the-chain-inside-a-link.md](the-chain-inside-a-link.md) ·\n[proving-a-pass-reliable.md](proving-a-pass-reliable.md) ·\n[earning-the-right-to-write.md](earning-the-right-to-write.md) ·\n[microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) ·\n[the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md) ·\n[writing-out-the-unwritten.md](writing-out-the-unwritten.md) ·\n[philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) ·\n[what-a-rule-rests-on.md](what-a-rule-rests-on.md) ·\n[kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md) ·\n[the-ratification-record-and-the-triage-feed.md](the-ratification-record-and-the-triage-feed.md) ·\n[corpus-graph/README.md](corpus-graph/README.md)\n"}